Metodologia para seleção de algoritmo de aprendizagem de máquina para estudos de controle centrado no usuário de edificações
Ir para navegação
Ir para pesquisar
Título principal
Metodologia para seleção de algoritmo de aprendizagem de máquina para estudos de controle centrado no usuário de edificações [recurso eletrônico] / Thayane Lodote Bilésimo ; orientador, Enedir Ghisi
Data de publicação
2024
Descrição física
130 p. : il.
Nota
Disponível somente em versão on-line.
Tese (doutorado) – Universidade Federal de Santa Catarina, Centro Tecnológico, Programa de Pós-Graduação em Engenharia Civil, Florianópolis, 2024.
Inclui referências.
Metodologia para seleção de algoritmo de aprendizagem de máquina para estudos de controle centrado no usuário de edificações [recurso eletrônico] / Thayane Lodote Bilésimo ; orientador, Enedir Ghisi
Data de publicação
2024
Descrição física
130 p. : il.
Nota
Disponível somente em versão on-line.
Tese (doutorado) – Universidade Federal de Santa Catarina, Centro Tecnológico, Programa de Pós-Graduação em Engenharia Civil, Florianópolis, 2024.
Inclui referências.
Abstract: The current building sector scenario demands new energy efficiency strategies. They need to be capable of identifying and assuring comfortable environments according to users’ perceptions. Machine learning – a method of data analysis capable of identifying patterns and make predictions – is an alternative to identify users’ preferences and control lighting and heating, ventilation and air-conditioning systems in buildings. Supervised learning is the most common approach because it allows classifying environmental conditions in terms of comfort. However, it is important to identify the algorithms’ capacity of making right predictions and adapting to the change in the preferences of building users over time. This research aims to present a method to assess different supervised learning algorithms. The method must allow the identification of the most accurate algorithms to be used in studies of occupant centered control systems. In order to assess algorithms’ performance in realistic scenarios, the following data were collected: occupancy, illuminance, temperature, and the current state of openings and artificial lighting system. Data were collected for eight months, between June 2022 and February 2023 in a research laboratory of Federal University of Santa Catarina. They were pre-processed and grouped in subsets within 1, 7, 15 and 30 days. Each subset was used in training and testing the following algorithms: Decision Tree, k-Nearest Neighbors, Multilayer Perceptron, Random Forest and Support Vector Machine, in order to predict the state of the lighting system. The results were statistically compared and the performance of the algorithms was assessed using each models’ accuracy, precision and recall. First, the ideal subset to each algorithm (in number of days) was identified. In the next step, algorithms were compared. The best configuration (algorithm and number of days) was submitted to an optimization process. Then, the update of the subset was assessed, aiming to evaluate the optimized algorithm working in a realistic scenario. For this case study, the best results were achieved using the k- Nearest Neighbors and subsets within 7 days for training. The algorithm was able to adapt to changes in users’ patterns and could reach a good performance even after total data substitution. At the end of the analysis, accuracy, precision and recall remained around 98%, in average. Finally, it is possible to affirm that the method proposed allowed to appropriately compare and select the algorithm and the ideal subset to predict the state of the artificial lighting system.